[Corpora-List] Coefficients in SVM models

Miles Osborne miles at inf.ed.ac.uk
Sat Aug 22 15:55:44 CEST 2015


I would add a note of caution here. Generally parameters are optimised in a global manner and it doesn't really make sense to interpret individual weights. For example, if every single example had an always-on feature which took a very large negative weight, other 'negative' weights might in fact become positive as they need to out weigh the negative weight.

I would suggest instead doing ablation studies (ie systematically removing sets of examples / features) as a much more direct and reliable way to see what is going on.

Miles

-- The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336. -------------- next part -------------- A non-text attachment was scrubbed... Name: not available Type: text/html Size: 807 bytes Desc: not available URL: <https://mailman.uib.no/public/corpora/attachments/20150822/dfc9279b/attachment.txt>



More information about the Corpora mailing list